Documentation

Agent skills

Guide series

Part 4 of 5 — Agent skills (you are here)

Previous: Agent memory  ·  Next: Building your agent

All guides — overview and full list

Skills are documents you write—usually in a dedicated workspace. Each skill is a short playbook: when to use it, what steps to follow, what to avoid. You can create as many skills as you want and reuse them across chats and workflows.

A1KnowHow injects skills into the agent context for you. Configure Skill disclosure on the agent in Settings → Agents (workspaces and/or pinned documents). Do not list skill document IDs in the system or user prompt, and do not instruct the model to search or read skills itself—that pattern is unreliable.

Optional YAML frontmatter on a skill document (name, description) helps the catalog and discovery matching. Document titles are yours to choose; there is no required title prefix.

Skills are documents you write

  1. Organize (left sidebar) → Create workspace if you need a dedicated skills area, or use an existing workspace. Open the workspace.
  2. Create Document for each playbook.
  3. Optionally add frontmatter at the top of the document:
---
name: remortgage-timing-window
description: Use when the user asks when to start planning a remortgage.
---
  1. Write the playbook body (When to use, Steps, Do not).
  2. In Settings → Agents → your agent → Skill disclosure, choose:
    • Pinned — always inject specific skill documents
    • Discovery — search configured skill workspaces using the user’s message
    • Auto — pinned documents if set, otherwise discovery
  3. Keep the System Prompt as role, output contract, and “follow injected skill blocks literally.” Keep the user prompt (chat or workflow step) as the task brief only.

See Building your agent and Creating agents.

What goes inside a skill

Keep each skill short (roughly one screen). A simple structure works well:

  • When to use — what kind of question triggers this playbook
  • Steps — numbered actions in order
  • Do not — mistakes to avoid (for example, do not quote live mortgage rates)
  • Examples (optional) — a sample user question

Example skill document

---
name: remortgage-timing-window
description: Use when the user asks when to start planning a remortgage.
---

# Remortgage timing window

## When to use

The user is roughly 3–12 months from the end of a fixed or tracker deal,
or asks when they should start planning a remortgage.

## Steps

1. Confirm the end date of their current deal (ask if you do not know it).
2. Explain the usual 3–6 month window: product transfer with current lender
   vs shopping with a new lender.
3. Mention early repayment charges if they might switch before the deal ends.
4. Suggest a simple timeline: when to start comparing, when to apply, when
   completion should target the deal end date (to avoid the standard variable rate).
5. If memory is configured, add the fix end date and a target "start shopping"
   date to the memory document.

## Do not

- Quote live interest rates or recommend a specific lender.
- Give regulated financial advice; frame options and suggest an FCA-regulated
  broker when appropriate.

## Examples

- "I'm 8 months from the end of my 2-year fix—help me plan."
- "When should I start looking for a new deal?"

You might add more skills over time (product types, fees checklist, documents to gather, and so on). The remortgage walkthrough in Building your agent uses two or three to start; you can grow the library later.

How skills reach the agent

When someone asks a question, A1KnowHow:

  1. Resolves the agent’s skill disclosure settings (pinned documents and/or skill workspaces).
  2. Injects a short catalog of available skills plus the selected skill bodies into the model context for that user message.
  3. Expects the agent to follow those injected instructions literally.

You do not need to tell the model to search the skills workspace or open document IDs. Configure skills on the agent; keep prompts focused on the job.

Skills are documents + agent settings

There is no separate “Skills” product outside documents. You write playbooks in a workspace, then attach that workspace or pin documents under Skill disclosure on the agent.

Prompt split (system vs user vs platform)

LayerWhat to put there
System promptRole, output contract, “follow injected skill blocks literally,” tool routing
User prompt / chat messageThe actual task, constraints, and template variables
PlatformSkill catalog + skill bodies from Skill disclosure

Listing skill document IDs in prompts and asking the model to read them is an anti-pattern—models often skip those reads.

Next

Continue to Building your agent for an end-to-end remortgage assistant that combines skills and memory.